To-do

[apcamargo]

• Use Vamb's transformation to reduce the number of TNF dimensions (103, instead of 136)
• Reduce memory footprint:
  • Use screed
  • Use hashes
  • Use Rust in taxopy
• Implement a modular interface, so that users can choose between several combinations. For example:
  • Sequence composition
  • Sequence composition + coverage
  • Sequence composition + coverage + codon usage
  • Sequence composition + coverage + codon usage + taxonomy
  • Coverage + codon usage
  • …

[jakobnissen]

If you need help implementing the transformation, let me know.
You can find a description of the idea in this paper, where I got the idea from: Kislyuk, A., Bhatnagar, S., Dushoff, J. et al. Unsupervised statistical clustering of environmental shotgun sequences. BMC Bioinformatics 10, 316 (2009). https://doi.org/10.1186/1471-2105-10-316

Practically speaking, I've found the best way is to compute a kernel beforehand using Scipy, save it to a file. then just load the kernel at runtime. You can just copy the file src/create_kernel.py directly from Vamb.

[apcamargo]

Thanks, @jakobnissen!

Now that I finished the first version of geNomad I might pick this up again. Your approach looks clean, I'll use it :)